/deeper_wider_siamese_trackers

This repo provides the source code and models of our deeper and wider siamese trackers. For CVPR2019 blind review.

Primary LanguagePythonMIT LicenseMIT

Deeper and Wider Siamese Networks for Real-Time Visual Tracking

This repo provides the source code and models of our deeper and wider siamese trackers for CVPR2019 blind review.

Introduction

SiamFC formulates the task of visual tracking as classification between background and target. SiamRPN improves SiamFC by introducing the robust region proposal estimation branch. However, the backbone network utilized in these trackers is still the classical AlexNet, which does not fully take advantage of the capability of modern deep neural networks.

Our proposals improve the performances of fully convolutional siamese trackers by,

  1. introducing CIR and CIR-D units to unveil the power of deeper and wider networks like ResNet and Inceptipon;
  2. designing reasonable backbones that are guilded by the analysis of how internal network factors (eg. receptive field, stride, output feature size) affect tracking performances.

Result snapshots

Requirements

Intel(R) Xeon(R) CPU E5-2630 v4 @ 2.20GHz GPU: NVIDIA GTX1080

  • python3.6
  • pytorch == 0.3.1
  • numpy == 1.12.1
  • opencv == 3.1.0

Important!! Since the speed of some opencv versions is relatively slow for some reason, it is recommended that you install packages above.

Tracking

SiamFC+ based on CIResNet-22 and CIR-Incep22 is provided for reproducing in this repo.

  • Tracking on a specific video
    • step1: place video frames and label file in dataset directory. David3 is provided in this demo.
    • step2: run,
python run_tracker.py --arch SiamFC_Res22 --resume ./pretrain/CIResNet22.pth --video David3 --vis True

--arch --resume video and vis indicate model name, pretrained model path, video name and visualization flag respectively.

  • Testing on VOT-2017 benchmark

    • step1: place VOT2017 benchmark files in dataset directory. The dataset directory should be organized as,
      |-- dataset
        |-- VOT2017
          |-- ants1
          |-- ...

    • step2: run,

      sh run_benchmark.sh SiamFC_Res22 ./pretrain/CIResNet22.pth VOT2017

    Screen snapshot shows like this (about 70 fps), and the results are saved in test directory.

    • step3 (optional): evaluate results with vot-toolkit. Please refer to offical vot-toolkit document for more information.

  • Testing on OTB-2013 benchmark

    • step1: place OTB2013 benchmark files in dataset directory.
    • step2: modify --dataset VOT2017 in run_benchmark.sh to --dataset OTB2013
    • setp3: run,
      sh run_benchmark.sh SiamFC_Res22 ./pretrain/CIResNet22.pth OTB2013
    • step4 (optional): evaluate results with otb-toolkit.

    If you want to test CIR-Incep22 model, simply modify step3 to,

    sh run_benchmark.sh SiamFC_Incep22 ./pretrain/CIRIncep22.pth OTB2013

  • Testing on other benchmarks

    If you want to test this demo on other benchmarks, please modify the code to your needs. Object tracking algorithms are sensitive to hyperparameters, so careful fine-tuneing for different benchmarks is necessary.


Other models will be released soon.

License

Licensed under an MIT license.